SparseCoding: A sparse coding hierarchy for realtime object recognition in complex scenes
FCT - Vision Laboratory - CINTAL/UAlg, EXPL/EEI-SII/1982/2013
Synopsis
This exploratory project as the main goal to combine two basic ideas:
(1) A kind of Hopfield neural network that is extended by an additional layer of neurons, called cliques, which code learned relations between basic patterns. The introduced redundancy leads to sparse codes and makes the new type of network very robust to incomplete data and noise. It is a form of neuromimetic computing, as a compact and fast model of our brain's neocortex which combines pattern recognition with associative memory.
(2) By building a hierarchy of such networks, translation- and rotation-invariant object recognition can be achieved. At a low level, very primitive and small patterns can be learned and detected, like edge fragments and corners linking edges. One level higher, these primitives are combined into bigger and more complex structures, with less dependency on precise localization. This is repeated until the top layer, where entire objects are coded irrespective of their position in the scene.
See also: link
Publications